Selective amnesia and system prompt risks in KV cache compression

On November 3, 2026, an article appeared on the Chinese developer platform Juejin examining a security question around KV Cache Compression: what happens when a model's KV cache is compressed during inference?

Rather than following the usual angle of cross-user sharing on shared infrastructure, the article asks a narrower question: if no cross-user sharing occurs, and the KV cache is compressed within a single model during a single inference pass, is that risk-free? The article argues that it is not. It describes two phenomena that can emerge from KV cache compression under such conditions. The first is selective amnesia, in which compression causes the model to drop or lose parts of its own context mid-inference. The second is system prompt leakage, in which material from the system prompt surfaces in ways that can end up in model output.

The piece positions itself as a technical reference for people working on inference optimization and safety design, framing these issues as hazards that fall outside the cross-user sharing scenario that typically dominates KV cache security discussions. It contains no benchmark numbers, no reproduced attacks, and no vendor statements.

As of the material available on this page, only this one public discussion has been published. No AI lab, cloud provider, or inference framework has responded to it, and no patch, paper, or rebuttal has appeared. The claims remain at the level of a single author's analysis rather than a documented incident, so the story currently rests on the article's framing alone.